Subzero: Add a peephole to fuse cmpxchg w/ later cmp+branch.

src/IceCfgNode.cpp

 //===- subzero/src/IceCfgNode.cpp - Basic block (node) implementation -----===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the CfgNode class, including the complexities
@@ skipping 86 matching lines @@
     Inst->updateVars(this);
   }
 }
 
 // This does part 2 of Phi lowering.  For each Phi instruction at each
 // out-edge, create a corresponding assignment instruction, and add
 // all the assignments near the end of this block.  They need to be
 // added before any branch instruction, and also if the block ends
 // with a compare instruction followed by a branch instruction that we
 // may want to fuse, it's better to insert the new assignments before
-// the compare instruction.
+// the compare instruction. The tryOptimizedCmpxchgCmpBr() method
+// assumes this ordering of instructions.
 //
 // Note that this transformation takes the Phi dest variables out of
 // SSA form, as there may be assignments to the dest variable in
 // multiple blocks.
 //
 // TODO: Defer this pass until after register allocation, then split
 // critical edges, add the assignments, and lower them.  This should
 // reduce the amount of shuffling at the end of each block.
 void CfgNode::placePhiStores() {
   // Find the insertion point.  TODO: After branch/compare fusing is
@@ skipping 366 matching lines @@
          I != E; ++I) {
       if (I != OutEdges.begin())
         Str << ", ";
       Str << "%" << (*I)->getName();
     }
     Str << "\n";
   }
 }
 
 } // end of namespace Ice